SAFE-MIL: a statistically interpretable framework for screening potential targeted therapy patients based on risk...

Area of Science:

• Oncology
• Computational Biology
• Biostatistics

Background:

• Targeted therapy offers significant benefits for patients with specific gene mutations.
• Inter-patient variability in mutation abundance leads to diverse survival outcomes, complicating risk assessment.
• Current models lack interpretability and rationality for predicting treatment failure.

Purpose of the Study:

• To develop a statistically interpretable framework for estimating treatment failure risk in targeted therapy.
• To address the challenge of varying survival benefits due to differences in mutation abundance.
• To provide a tool for accurate patient stratification in personalized medicine.

Main Methods:

• Developed SAFE-MIL, a framework integrating multi-instance learning (MIL) with the Hosmer-Lemeshow test.
• Constructed patient effectiveness labels and sampled patients into groups using MIL.
• Designed a novel interpretable loss function based on the Hosmer-Lemeshow test for risk estimation.

Main Results:

• SAFE-MIL accurately estimates drug treatment failure risk and provides optimal risk stratification thresholds.
• In a case study of 457 non-small cell lung cancer patients, SAFE-MIL outperformed traditional regression methods in accuracy.
• The framework effectively captures inter-patient variability in risk, offering statistical interpretability.

Conclusions:

• SAFE-MIL provides an interpretable computational framework for risk assessment in targeted therapy.
• The model accurately guides clinical decision-making for drug use and patient stratification.
• SAFE-MIL enhances precision in personalized medicine and is applicable to other patient stratification problems.